EDA for Affective AI datasets

Imports

KDEF Dataset

The dataset files are named using a specific convention that encodes information about the subject, expression, and angle of the photograph. The naming convention is as follows:

Example: AF01ANFL.JPG

• Letter 1: Session
  • A = series one
  • B = series two
• Letter 2: Gender
  • F = female
  • M = male
• Letter 3 & 4: Identity number
  • 01 - 35
• Letter 5 & 6: Expression
  • AF = afraid
  • AN = angry
  • DI = disgusted
  • HA = happy
  • NE = neutral
  • SA = sad
  • SU = surprised
• Letter 7 & 8: Angle
  • FL = full left profile
  • HL = half left profile
  • S = straight
  • HR = half right profile
  • FR = full right profile
• Extension: Picture format
  • JPG = jpeg (Joint Photographic Experts Group)

Dataset Statistics

• Participants: 70 (35 males and 35 females).
• Age: m 25 years, ranging from 20 to 30 years.
• Expressions: 7 (neutral, happy, angry, afraid, disgusted, sad, surprised).
• Angles: 5 (full left profile, half left profile, straight, half right profile, full right profile).
• Sessions: 2.
• Number of pictures: 4900.
• Size: 562 * 762 pixels.
• Resolution: 72*72 dpi.
• Colors: 16.7 million (32 bit).
• Size inflated: 1.6 Mb.
• Size compressed: approximately 122 kb (ranging from 85 to 158 kb).
• File format: JPEG.
• Compression quality: 94 %.

Check if files are valid:

For each directory (person) check:

• If every person has the same number of images (35)

  For each image check:

  • If it's JPG
  • If it starts with the directory name
  • If the emotion code is valid
  • If the angle code is valid
  • If every image is of size 562x762
  • If is RGB

• Count number of images per emotion and angle, they should be 7 and 5 respectively

Failures:

• 2 instances failed on wrong naming convention: "AF31V.JPG", "AM31H.JPG" (fix: renamed)

• 2 instance failed on image size: "AM05ANFR.JPG", "BM10ANFR.JPG" - size 970 x 633 (fix: manually cropped)

• 10 instances failed on image being mostly black (mean pixel value < 10):

  • "AF01SUFR.JPG", "AF10AFFR.JPG", "AF11NEHL.JPG", "AF20DIHL.JPG", "AM25DIFL.JPG", "AM34DIFR.JPG", "BF13NEHR.JPG", "BM21DIFL.JPG", "BM22DIHL.JPG", "BM24DIFL.JPG" (fix: TODO)

Example of single person for the same expression with different angles:

Example of how different people show emotions:

Original vs Grayscaled vs Degraded

Histogram analysis

Average emotion images: